Method for advanced prediction of electric fire situation of long-distance tunnel moving train based on PINN

Abstract

The application discloses a long and large interval tunnel moving train electrical fire situation advanced prediction method based on PINN, which comprises the following steps: 1, a long and large interval tunnel moving train fire experiment platform is built, and a physical information dataset of long and large interval tunnel moving train electrical fire is constructed; 2, a long and large interval tunnel moving train electrical fire situation advanced prediction neural network based on PINN is constructed; 3, a loss function set of the long and large interval tunnel moving train electrical fire advanced prediction neural network is constructed; 4, the loss function set is trained to obtain an optimal situation prediction model, and the long and large interval tunnel moving train electrical fire situation advanced prediction is realized. According to the multi-modal physical information of the long and large interval tunnel moving train electrical fire in the current period, the long and large interval tunnel moving train electrical fire situation advanced prediction neural network based on PINN can realize the advanced prediction of the fire disaster field parameters of the moving train under the coupling of the complex motion law and the tunnel boundary condition, and the prediction accuracy is high.

Description

Technical Field

[0001] This invention belongs to the field of fire safety and intelligent sensing technology in tunnels and underground spaces. Specifically, it relates to an advanced prediction method for electrical fire situations of moving trains in long tunnels based on PINN, which is applicable to the situation prediction and model deduction of electrical fires of trains in long tunnels such as intercity tunnels and railway tunnels. Background Technology

[0002] Subway tunnels have become an important part of urban rail transit networks. Unlike general surface transportation facilities, subway tunnels are characterized by their length, strong enclosure, and difficulties in evacuation and rescue. However, modern subway trains have a large number of electrical equipment such as pantographs, distribution cabinets, and air conditioning units, which are prone to electrical fires due to overload, aging lines, and environmental interference.

[0003] When a train catches fire during operation, regulations require it to continue to the next rescue station for assistance, a process that may extend through the entire length of a long tunnel. If the train is on fire within the tunnel, the high-temperature smoke is difficult to expel quickly due to the confined space, causing the fire to spread rapidly and posing a serious threat to passenger safety and the stability of the train structure. Furthermore, the combustion characteristics of this type of fire are influenced by multiple factors, including train speed and tunnel airflow, resulting in complex spatiotemporal distributions of temperature and heat flux density that traditional models struggle to accurately predict. Therefore, achieving rapid, accurate, and proactive prediction of the thermal disaster situation of trains in long tunnels with complex motion patterns is one of the core technical challenges that urgently needs to be addressed in the field of tunnel and underground space fire safety.

[0004] Currently, monitoring of the electrical fire and thermal disaster situation (including temperature field and heat flow distribution) of moving trains in long tunnels mainly relies on point-by-point sampling by sensors deployed on the train and tunnel ceiling. This method usually has the following shortcomings: (1) The number of sensors deployed is limited, the means of fire source detection are limited, and the acquired thermal disaster field data are discretely distributed, which cannot reflect the continuous spatiotemporal evolution characteristics of the thermal disaster field situation. (2) Existing monitoring methods are mostly real-time monitoring technologies, which cannot achieve advanced prediction and are difficult to achieve early warning of train electrical fires and prediction of personnel evacuation. They are also poorly applicable to the prediction and early warning of electrical fires of moving trains in long tunnels. (3) Current prediction methods mostly rely on calculations based on semi-empirical formulas or single physical models, which makes it difficult to predict electrical fires on moving trains with complex coupling factors, resulting in a large deviation between the prediction results and the actual fire behavior.

[0005] Therefore, there is an urgent need to develop a method for predicting fire situations in advance, so as to achieve the prediction of electrical fire situations of trains moving in long tunnels under complex motion patterns, thereby providing reliable technical support for fire safety prevention and control in rail transit. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes an advanced prediction method for electrical fire situations of moving trains in long-distance tunnels based on PINN. This method enables advanced prediction of temperature and heat flux density in electrical fires of moving trains in long-distance tunnels, meeting the urgent need for fire situation simulation in urban rail transit and thereby improving the safety and reliability of emergency response to electrical fires in long-distance tunnels.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN, which includes the following steps; Step 1: Collect and construct an electrical fire dataset for long-distance tunnel moving trains using an experimental platform, including: a physical boundary dataset for long-distance tunnel moving trains. A collection of transient flame images of trains moving through long tunnels. Thermal disaster field dataset of moving trains in long tunnel sections ,in, , , Indicates the first i Physical boundary datasets, transient flame image datasets, and thermal disaster field datasets under various operating conditions. N This indicates the total number of operating conditions; and , Indicates the first i Under the first working condition m Physical boundary parameters at each sampling time, Indicates the first i The total number of data collection times under each working condition; and ,in, Indicates the first i Under the first working condition m Data set of thermal disaster fields at each sampling time, , , They represent the first i Under the first working condition m The vector of experimental values ​​for roof temperature, radiative heat flux density, and convective heat flux density at each sampling time, and , , ,in, , , The first i Under the first working condition The sampling time of the first sampling moment j The temperature, the first j The radiative heat flux density and the first j Experimental values ​​of convective heat flux density at each measuring point. For the first i The total number of measuring points deployed under each working condition; right Perform fragmentation and preprocessing, the first i Atlas of Probability Cloud Map of Flame Output under Various Operating Conditions and its flame size dataset ,in, Indicates the first i Under the first working condition m Flame probability cloud map corresponding to each sampling time. Indicates the first i Under the first working condition m Flame size parameters corresponding to each sampling time; H and W for Height and width, for The number of channels; Step 2: Construct a train electrical fire prediction network, including: a physical information extraction module, a SwinTransformer flame feature extraction module, an FNO temperature / heat flux field extraction module, a multimodal information fusion module, and a temperature / heat flux density prediction head; for , , and the i Under the first working condition The moment to the Time series of thermal disaster fields within a given moment Process and output the first... i Under the first working condition The moment to the The predicted value sequence of thermal disaster fields within a certain time period ,in ,in, For the time span of the predicted value sequence, , , They represent the first i Under the first working condition The predicted vector of temperature, radiative heat flux density, and convective heat flux density at each sampling time, and , , ,in, , , The first i Under the first working condition The sampling time of the first sampling moment j The temperature, the first j The radiative heat flux density and the first j Predicted values ​​for each convective heat flux density measurement point; Step 2.1, the physical information extraction module extracts information through the first fully connected layer. After performing the dimensionality upgrade operation, we obtain the first... i Under the first working condition m physical information parameter feature vector at each sampling time Then, the ReLU activation function is applied to... Perform a nonlinear transformation to obtain the first... i Under the first working condition m The eigenvectors of nonlinear parameters at each sampling time point , The input is then fed into the second fully connected layer for processing, resulting in the... i Under the first working condition m Physical feature vector at each sampling time ; Step 2.2, the Swin Transformer flame feature extraction module, includes: an image segmentation and linear embedding layer, a Swin Transformer encoder, a global feature aggregation head, and a flame geometry mapping layer, and performs... and Processing is performed to obtain the first... i Under the first working condition m Flame feature vector at each sampling time ; Step 2.3, the FNO temperature / thermal flow field extraction module includes: a linear projection layer, a temporal context stacking module, Layer FNO operator block, feature output module, and for Perform time series feature extraction, and the output includes the first... i Under the first working condition The moment to the The fusion feature vector of thermal disaster field information within a given time period ; Step 2.4, , and Perform fusion and splicing to obtain the first... i Under the first working condition The moment to the The multimodal fusion vector within each time step ; Step 2.5, The input temperature / heat flux density is processed in the advanced prediction head to obtain the final result. ; Step 3, based on the first i Under the first working condition From the sampling time to the... The predicted and experimental values ​​within each sampling time point are used to construct a train electrical fire prediction network at the [number]th sampling time point. i Total loss function under each working condition Including: temperature loss function Radiative heat flux density loss function convective heat flux density loss function ; Step 4: Iteratively train the train electrical fire prediction network using the gradient descent algorithm and calculate the total loss function under all operating conditions. Adjust the network parameters until the total loss function is achieved. The process continues until convergence, resulting in a trained neural network model for predicting electrical fire situations. This model is then used to detect electrical fires from the input dataset and outputs the optimal prediction sequence.

[0008] The present invention also features a method for predicting electrical fire situations of moving trains in long tunnels based on PINN, wherein step 2.2 includes: Step 2.2.1, the image segmentation and linear embedding layer includes: an image segmentation layer and a linear embedding layer, and is applied sequentially to... Processing is performed to obtain the first... i Under the first working condition m The initial input feature tensor at each sampling time ,in, C for The number of channels; Step 2.2.2: The Swing Transformer encoder contains four Swing Transformer base blocks and an image block merging layer between two adjacent Swing Transformer base blocks. Each Swin Transformer basic block includes, in sequence: layer normalization module, window multi-head self-attention module, feedforward neural network module, and moving window multi-head self-attention module; After four stages of feature processing, the output is... i Under the first working condition m The final output feature tensor at each sampling time point in stage 4 ; Step 2.2.3, Global Feature Aggregation Head Pair Perform global average pooling to obtain the first... iUnder the first working condition m Global visual feature vector at each sampling time point ; Step 2.2.4, the flame geometry mapping layer is processed by MLP. Encode to obtain the first i Under the first working condition m Flame size feature vector at each sampling time Finally, and Vector concatenation is performed, and a non-linear mapping is applied through an MLP to finally output the first vector. i Under the first working condition m Flame feature vector at each sampling time .

[0009] Furthermore, step 2.2.2 includes: Step 2.2.2.1, when s When =1, let For the first i Under the first working condition m The sampling time at the ... s Input feature tensor of the stage ; Step 2.2.2.2, will Enter the first s Within the first Swing Transformer base block, and after processing by the layer normalization module, the second... s The input feature tensor after each stage of normalization ; Step 2.2.2.3, Enter the first s The window multi-head self-attention module of the basic Swing Transformer block, and the window multi-head self-attention module will... Divide into several sizes Non-overlapping local windows; Calculate the first value in any non-overlapping local window p Embedded vectors The corresponding multi-head self-attention, yielding the first p Multi-head attention fusion feature vector under each query vector ,right Perform an index rearrangement operation to obtain the first... i Under the first working condition m The sampling time at the ... s Multi-head attention fusion feature tensor of the multi-head self-attention module of the window at each stage ; then Perform a linear transformation to obtain the first... i Under the first working conditionm The sampling time at the ... s Output feature tensor of the multi-head self-attention module in the phased window ; Step 2.2.2.4, the first s The feedforward neural network module in the basic Swing Transformer block uses equation (1) to... Processing is performed to obtain the first... i Under the first working condition m The sampling time at the ... s intermediate feature tensor of the stage ; (1) In equation (1), + indicates residual connection; Presentation layer normalization operation; Step 2.2.2.5, the first s The multi-head self-attention module of the moving window in the basic Swing Transformer block utilizes equation (2) to... Perform window offset and output the first... i Under the first working condition m The sampling time at the ... s Feature tensor after phase shift ; (2) In equation (2), Roll represents a cyclic shift operation. This indicates the floor function; the negative sign indicates a shift along the upper left direction. Indicates the size of the non-overlapping local window; Step 2.2.2.6, will Divide into several sizes The offset window, and calculate the first offset window within any offset window. p Embedded vectors The corresponding multi-head self-attention features are then processed in the same manner as in step 2.2.2.3. , obtained the i Under the first working condition m The sampling time at the ... s Multi-head self-attention fusion feature tensor of the multi-head self-attention module of the moving window at each stage ; Step 2.2.2.7, using equation (3) to... Perform a reverse circular shift to obtain the first... i Under the first working condition m The sampling time at the ... s Output feature tensor of the multi-head self-attention module in the phased moving window ; (3) In equation (3), This indicates a reverse cyclic shift operation; Step 2.2.2.8, using equation (1) to... Processing is performed to obtain the first... i Under the first working condition m The sampling time at the ... s Stage output feature tensor ; Step 2.2.2.9, Enter the first s In the image patch merging layer, and Merging adjacent 2×2 image blocks in the middle yields the first... i Under the first working condition m The sampling time at the ... s Feature tensor after stage merging And perform a linear transformation to obtain the first... i Under the first working condition m The sampling time at the ... s +1 stage input feature tensor ; Step 2.2.2.10, will Assign to s Then, return to step 2.2.2.1 and execute sequentially until the output of the 4th SwingTransformer base block is obtained. .

[0010] Furthermore, step 2.3 includes: Step 2.3.1, the linear projection layer passes through an MLP head pair Perform a dimensionality upgrade operation to obtain the first... i Under the first working condition m Embedded feature matrix at each sampling time point Thus, the first i Under the first working condition From the sampling time to the... Embedded feature sequence within each sampling time point ; Step 2.3.2, the timing context stacking module will... Stack each element in the array and output the first element. i Under the first working condition From the sampling time to the... The fused tensor within each sampling time point ; Will After reorganizing along the time dimension, we obtain the first... i Under the first working condition From the sampling time to the... The initial input feature matrix of time series information within each sampling time point ; Step 2.3.3 Layer FNO operator block uses equation (4) for After processing, the final result is obtained. i Under the first working condition From the sampling time to the... The time sequence information within the sampling time period of the first sampling moment Layer FNO operator block output feature matrix ; (4) In equation (4), Indicates the first The output feature matrix of the -1 layer FNO operator block Indicates the first The complex weights to be learned in the layer FNO operator block. Indicates the inverse Fourier transform. Represents the Discrete Fourier Transform. Indicates a skip connection; Indicates the first The output feature matrix of the layer FNO operator block; when season ; Step 2.3.4: The feature output module first passes the feature through an MLP layer. Perform feature mapping to obtain the first... i The first working condition includes the From the sampling time to the... High-dimensional hidden layer feature matrix of temporal information within each sampling time point Subsequently, regarding Perform pooling operation to obtain .

[0011] Furthermore, step 3 includes: Step 3.1, construct the first equation from equation (5). i Temperature loss function under various operating conditions ; (5) In equation (5), For the first i Data loss item for temperature under each operating condition. For the first i Physical loss due to temperature under various operating conditions. These are the weighting coefficients for the temperature loss function; Step 3.2, construct the first equation from equation (6). iLoss function of radiative heat flux density under various operating conditions ; (6) In equation (6), For the first i Data loss term for radiative heat flux density under each operating condition. For the first i The physical loss term of radiative heat flux density under each operating condition. The weighting coefficients for the radiative heat flux density loss function; Step 3.3, construct the first equation from equation (7). i Loss function of convective heat flux density under various operating conditions ; (7) In equation (7), For the first i Data loss term for convective heat flux density under each operating condition. For the first i Physical loss term of convective heat flux density under each operating condition. The weighting coefficients for the convective heat flux density loss function; Step 3.4, construct the first using equation (8). i Total loss function under each working condition : (8).

[0012] Furthermore, step 3.1 includes: Step 3.1.1, construct using equation (9) ; (9) Step 3.1.2: Based on the PDE residuals, initial condition residuals, and boundary condition residuals, construct using equation (10). ; (10) In equation (10), For the first i Under the first working condition u PDE residuals at each collocation point temperature For the first i Under the first working condition u The initial condition residuals of temperature at each collocation point. For the first i Under the first working condition u Boundary condition residuals for temperature at each collocation point; , and These represent the total number of collocation points for the partial differential equation, initial conditions, and boundary conditions, respectively.

[0013] Furthermore, step 3.2 includes: Step 3.2.1, construct using equation (11) ; (11) Step 3.2.2: Based on the PDE residuals, initial condition residuals, and boundary condition residuals, construct using equation (12). ; (12) In equation (12), For the first i Under the first working condition u PDE residuals of radiative heat flux density at each collocation point For the first i Under the first working condition u Initial condition residuals of radiative heat flux density at each collocation point For the first i Under the first working condition u Boundary condition residuals of radiative heat flux density at each collocation point.

[0014] Furthermore, step 3.3 includes: Step 3.3.1, construct using equation (13) ; (13) Step 3.3.2: Based on the PDE residuals, initial condition residuals, and boundary condition residuals, construct using equation (14). ; (14) In equation (14), For the first i Under the first working condition u PDE residuals of convective heat flux density at each collocation point For the first i Under the first working condition u Initial condition residuals of convective heat flux density at each collocation point For the first i Under the first working condition u Boundary condition residuals of convective heat flux density at each collocation point.

[0015] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program that supports the processor in executing a method for predicting the electrical fire situation of a moving train in a long tunnel, and the processor is configured to execute the program stored in the memory.

[0016] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of a method for predicting the electrical fire situation of a moving train in a long tunnel.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a forward prediction neural network based on PINN, which can make forward predictions of the electrical fire and thermal disaster field situation of moving trains in long tunnels based on the multimodal physical information of the current time period. This breaks through the limitations of traditional real-time monitoring methods and significantly improves the timeliness of urban rail transit fire prevention and control.

[0018] 2. This invention constructs a multimodal physical information fusion prediction network based on PINN. By embedding the fire dynamics control equation into the loss function, the model has strong generalization ability and effectively overcomes the defects of traditional pure data-driven models that are prone to overfitting and have poor extrapolation ability. At the same time, by deeply fusing tunnel physical information, flame images and thermal disaster field features through multimodal processing, the robustness of the model in complex fire scenarios is significantly improved.

[0019] 3. This invention uses a Fourier feature mapping network to encode the input time-series thermal disaster field data. Compared with traditional recurrent neural network architectures such as long short-term memory networks, it can map low-dimensional time-series inputs to high-dimensional space, effectively capturing multi-scale dynamic features in the fire evolution process. At the same time, it avoids the error accumulation and gradient vanishing problems caused by recursive calculations, further improving the accuracy and stability of advance prediction.

[0020] 4. This invention constructs a fire dataset covering multiple motion conditions by building a scaled-down model based on the Froude similarity principle, and uses the PINN neural network to achieve end-to-end multi-parameter advance prediction. The entire process does not require manual feature extraction or parameter adjustment. The network can adapt to multi-factor coupled dynamic fire scenarios under different train motion conditions, effectively reducing the reliance on manual operation and providing technical support for the digital and intelligent development of urban rail transit safety management. Attached Figure Description

[0021] Figure 1 This is a flowchart of the fire prediction network for moving trains in long tunnels according to the present invention. Figure 2 This is a diagram of the Swing Transformer network architecture of the present invention; Figure 3 This is a structural diagram of the Fourier Neural Operator of this invention; Figure 4 This is a diagram of the experimental device for a long-distance tunnel moving train according to the present invention; The following components are labeled in the diagram: 1. Meshing drive system, 2. Optical tracking system, 3. Action camera, 4. Train guide rail, 5. Air supply system, 6. Train model, 7. Porous burner, 8. Tunnel model, 9. Circulating water pump, 10. Portable data acquisition module, 11. Servo motor. Detailed Implementation

[0022] In this embodiment, to address the shortcomings of traditional data prediction models when facing fires involving pantographs on the roofs of high-speed trains in long tunnels—namely, difficulty in fully utilizing multimodal fire field information, low prediction accuracy under complex conditions such as train speed changes, and poor generalization ability—a method for predicting electrical fire situations on trains in long tunnels based on PINN is proposed. This method utilizes a dual-path network of Swing Transformer and FNO Fourier neural operators to extract morphological features of transient flame images and temporal features of thermal disaster field data. Combined with PINN loss term constraint network training, fire dynamics are embedded into the data-driven model, thereby improving the model's generalization ability and prediction accuracy under complex conditions. This enables advanced prediction of thermal disaster field data for future periods of pantograph fires on train roofs, providing a basis for automatic alarm and early response to train fires. Figure 1 As shown, the specific process includes the following steps: Step 1: Construct an experimental platform for electrical fire simulation of high-speed trains operating in long tunnels. For detailed platform construction, refer to the "Fire Simulation Device and Application for High-Speed ​​Trains Operating in Long Tunnels" (CN116052501B). Figure 4 As shown, the dataset for collecting and constructing electrical fire data for moving trains in long tunnels includes: a dataset of the physical boundaries of moving trains in long tunnels. A collection of transient flame images of trains moving through long tunnels. Thermal disaster field dataset of moving trains in long tunnel sections ,in, , , Indicates the first i Physical boundary datasets, transient flame image datasets, and thermal disaster field datasets under various operating conditions. N This indicates the total number of operating conditions; and , Indicates the first i Under the first working condition m Physical boundary parameters at each sampling time, Indicates the first i The total number of data collection times under each working condition; and ,in, Indicates the first i Under the first working condition m Data set of thermal disaster fields at each sampling time, , , They represent the first i Under the first working condition m The vector of experimental values ​​for roof temperature, radiative heat flux density, and convective heat flux density at each sampling time, and , , ,in, , , The first i Under the first working condition The sampling time of the first sampling moment j The temperature, the first j The radiative heat flux density and the first j Experimental values ​​of convective heat flux density at each measuring point. For the first i The total number of measuring points deployed under each working condition.

[0023] For train fires in tunnel sections, flame image information is easily affected by tunnel ambient light, smoke obstruction, and motion vibration. In this embodiment, to remove background noise from the flame image and improve the effective signal-to-noise ratio, [the following measures are taken]. To perform segmentation and preprocessing operations, the sampling time is first... m The time range of 0.5 seconds before and after is called the time window. w Based on time window w Will Divide into several segments, denoted as , Indicates the first i Under various working conditions w A set of flame images within a time window.

[0024] Using the Otsu algorithm, based on equation (1) Each frame of the flame image is binarized to obtain the first frame. i Under various working conditions w Binarized flame image set within the time window ; (1) In equation (1), This represents the optimal segmentation threshold determined by the Otsu algorithm. , Represents image set , The Middle k Frame image located at pixel coordinates The Otsu algorithm determines the optimal threshold by maximizing the inter-class variance between the flame region and the background region. This process can reduce the impact of noise such as tunnel background, ambient light, and fire smoke on the flame image.

[0025] Based on equation (2) The binary image in the image is averaged over time to obtain the first... i Under various working conditions w Flame probability cloud map within the time window Considering w and m One-to-one correspondence, Can write , Indicates the first i Under the first working condition m The flame probability cloud map corresponding to each sampling time point; the pixel size of the cloud map in this embodiment. Take after processing The number of channels is Take 1; (2) In equation (2), K express The number of frames in the flame image, in this embodiment, is taken as... K =30.

[0026] exist Extract the 0.5 probability contour lines. and based on Extract the first i Under the first working condition m Flame size parameters at each sampling time .

[0027] After completing the above fragmentation and preprocessing, the output is the first... i Atlas of Probability Cloud Map of Flame Output under Various Operating Conditions and its flame size dataset .

[0028] Step 2: Construct a train electrical fire prediction network, including: a physical information extraction module, a SwinTransformer flame feature extraction module, an FNO temperature / heat flux field extraction module, a multimodal information fusion module, and a temperature / heat flux density prediction head; for , , and the i Under the first working condition The moment to the Time series sequence of thermal disaster fields within a given moment Process and output the first... i Under the first working condition The moment to the The predicted value sequence of thermal disaster fields within a certain time period ,in ,in, For the time span of the predicted value sequence, , , They represent the first i Under the first working condition The predicted vector of temperature, radiative heat flux density, and convective heat flux density at each sampling time, and , , ,in, , , The first i Under the first working condition The sampling time of the first sampling moment j The temperature, the first j The radiative heat flux density and the first j The predicted values ​​for each convective heat flux density measurement point; in this embodiment; The time frame is 5 seconds, which means using multimodal information within 5 consecutive seconds to predict the changes in roof temperature and heat flux density within the next 5 seconds.

[0029] Step 2.1, the physical information extraction module extracts information through the first fully connected layer. After performing the dimensionality upgrade operation, we obtain the first... i Under the first working condition m physical information parameter feature vector at each sampling time Then, the ReLU activation function is applied to... Perform a nonlinear transformation to obtain the first... i Under the first working condition m The eigenvectors of nonlinear parameters at each sampling time point , The input is then fed into the second fully connected layer for processing, resulting in the... i Under the first working condition m Physical feature vector at each sampling time .

[0030] Step 2.2: The Swin Transformer utilizes a multi-head self-attention mechanism to capture morphological information in key areas such as the flame core and edges, and achieves cross-regional information interaction through a shifted window approach. This solves the problems of high computational cost and lack of local inductive bias in the traditional VisionTransformer's global self-attention mechanism, reducing computational cost while simultaneously extracting local details and global structural information of the flame morphology. The Swin Transformer flame feature extraction module includes: an image segmentation and linear embedding layer, a Swin Transformer encoder, a global feature aggregation head, and a flame geometry mapping layer, and performs... and Processing is performed to obtain the first... i Under the first working condition m Flame feature vector at each sampling time .

[0031] Step 2.2.1, the image segmentation and linear embedding layer includes: an image segmentation layer and a linear embedding layer; firstly, the input... According to size The non-overlapping regions are divided to obtain... In this embodiment, there are several image blocks. P The value is 4, obtained after image segmentation. The nth image patch is flattened into a one-dimensional vector and linearly mapped to obtain the nth image patch. i Under the first working condition m The flame image corresponding to the sampling time is the first b Embedding vectors of image patches Then N indivual Reorganizing along spatial dimensions yields the first... i Under the first working condition m The initial input feature tensor at each sampling time ,in, C for The number of channels.

[0032] Step 2.2.2: The Swing Transformer encoder consists of four Swing Transformer base blocks and an image block merging layer between adjacent Swing Transformer base blocks. The specific architecture is as follows: Figure 2 As shown; Each Swin Transformer basic block includes, in sequence: layer normalization module, window multi-head self-attention module, feedforward neural network module, and moving window multi-head self-attention module.

[0033] After four stages of feature processing, the output is... i Under the first working condition m The final output feature tensor at each sampling time point in stage 4 In this embodiment, the embedding feature dimensions of the four stages are 96, 192, 384 and 768 respectively, the number of basic blocks in each stage are 2, 2, 2 and 2 respectively, the number of multi-head attention heads are 3, 6, 12 and 24 respectively, and the dimension of the hidden layer of the feedforward neural network is 4 times the embedding feature dimension of the corresponding stage.

[0034] Step 2.2.2.1, when s When =1, let For the first i Under the first working condition m The sampling time at the ... s Input feature tensor of the stage ; Step 2.2.2.2, will Enter the first s Within the first Swing Transformer base block, and after processing by the layer normalization module, the second... s The input feature tensor after each stage of normalization .

[0035] Step 2.2.2.3, Enter the first s The window multi-head self-attention module of the basic Swing Transformer block, and the window multi-head self-attention module will... Divide into several sizes Non-overlapping local windows; in the multi-head self-attention module, using local windows to calculate self-attention can limit the computational range of a single attention layer, reduce the computational load caused by global attention, improve the prediction efficiency of the network, and at the same time preserve the local geometric information of the flame outline, flame core, and edge regions to ensure the accuracy of the prediction; in this embodiment, the size of the non-overlapping local window is taken as... .

[0036] Calculate the first value in any non-overlapping local window p Embedded vectors The corresponding multi-head self-attention, yielding the first p Multi-head attention fusion feature vector under each query vector ,right Perform an index rearrangement operation to obtain the first... i Under the first working condition m The sampling time at the ... s Multi-head attention fusion feature tensor of the multi-head self-attention module of the window at each stage ; then Perform a linear transformation to obtain the first... i Under the first working condition m The sampling time at the ... s Output feature tensor of the multi-head self-attention module in the phased window .

[0037] Step 2.2.2.4, the first s The feedforward neural network module in the basic Swing Transformer block uses equation (1) to... Processing is performed to obtain the first... i Under the first working condition m The sampling time at the ... s intermediate feature tensor of the stage ; (3) In equation (3), + indicates residual connection; Presentation layer normalization operation.

[0038] Step 2.2.2.5: The moving window multi-head self-attention module shifts the window boundaries by half the size of the window, allowing the image block at the boundary of the previous window to enter the new attention window, thereby establishing information connections between adjacent windows. This enables cross-regional feature transfer between the flame core region, the flame edge region, and the adjacent background region, making up for the deficiency of window multi-head self-attention which only focuses on features within the window.

[0039] No. s The multi-head self-attention module of the moving window in the basic Swing Transformer block utilizes equation (4) to... Perform window offset and output the first... i Under the first working condition m The sampling time at the ... s Feature tensor after phase shift ; (4) In equation (4), Roll represents a cyclic shift operation. This indicates the floor function; the negative sign indicates a shift along the upper left direction. Indicates the size of the non-overlapping local window.

[0040] Step 2.2.2.6, will Divide into several sizes The offset window, and calculate the first offset window within any offset window. p Embedded vectors The corresponding multi-head self-attention features are then processed in the same manner as in step 2.2.2.3. , obtained the i Under the first working conditionm The sampling time at the ... s Multi-head self-attention fusion feature tensor of the multi-head self-attention module of the moving window at each stage .

[0041] Step 2.2.2.7, using equation (5) to... Perform a reverse circular shift to obtain the first... i Under the first working condition m The sampling time at the ... s Output feature tensor of the multi-head self-attention module in the phased moving window ; (5) In equation (5), This indicates a reverse cyclic shift operation.

[0042] Step 2.2.2.8, using equation (1) to... Processing is performed to obtain the first... i Under the first working condition m The sampling time at the ... s Stage output feature tensor .

[0043] Step 2.2.2.9, Enter the first s In the image patch merging layer, and The adjacent 2×2 image blocks are merged to obtain the first... i Under the first working condition m The sampling time at the ... s Feature tensor after stage merging And perform a linear transformation to obtain the first... i Under the first working condition m The sampling time at the ... s +1 stage input feature tensor In this embodiment, the image block merging layer at each stage will merge adjacent... Image patches are stitched together along the channel dimension to obtain intermediate features with four times the number of channels. Then, a linear layer reduces the number of channels to twice the original number. Through the processing of the image merging layer, the four stages sequentially capture local flame features to global flame features. After the four stages of processing, the number of feature channels increases from 96 to 768.

[0044] Step 2.2.2.10, will Assign to s Then, return to step 2.2.2.1 and execute sequentially until the output of the 4th SwingTransformer base block is obtained. .

[0045] Step 2.2.3, Global Feature Aggregation Head Pair Perform global average pooling to obtain the first... i Under the first working condition m Global visual feature vector at each sampling time point .

[0046] Step 2.3.1, the linear projection layer passes through an MLP head pair Perform a dimensionality upgrade operation to obtain the first... i Under the first working condition m Embedded feature matrix at each sampling time point Thus, the first i Under the first working condition From the sampling time to the... Embedded feature sequence within each sampling time point .

[0047] Step 2.3.2, the timing context stacking module will... Stack each element in the array and output the first element. i Under the first working condition From the sampling time to the... The fused tensor within each sampling time point ; Will After reorganizing along the time dimension, we obtain the first... i Under the first working condition From the sampling time to the... The initial input feature matrix of time series information within each sampling time point .

[0048] Step 2.3.3: In this embodiment, the number of FNO operator block layers is 4, the number of spectral truncation modes is 16, and the activation function of each layer is the GELU function. Layer FNO operator block uses equation (6) for Process, such as Figure 3 As shown, specifically, each layer of FNO operator blocks first processes the first... i Under the first working condition m The sampling time at the ... l Input feature matrix of layer FNO operator block Perform a discrete Fourier transform, and then pass the first... l Learnable complex weights of layers Adaptive adjustment is performed on each frequency component, and finally, an inverse Fourier transform is applied to the result to obtain the th frequency component. i Under the first working condition From the sampling time to the... Within the sampling time period, the first l Output characteristics of layer FNO operator blocks Finally, after processing through 4 layers of FNO operator blocks, the result is obtained. i Under the first working condition From the sampling time to the... The time sequence information within the sampling time point Layer FNO operator block output feature matrix ; (6) In equation (6), Indicates the first The output feature matrix of the -1 layer FNO operator block Indicates the first The complex weights to be learned in the layer FNO operator block. Indicates the inverse Fourier transform. Represents the Discrete Fourier Transform. Indicates a skip connection; Indicates the first The output feature matrix of the layer FNO operator block; when season .

[0049] Step 2.3.4: The feature output module first passes the feature through an MLP layer. Perform feature mapping to obtain the first... i The first working condition includes the From the sampling time to the... High-dimensional hidden layer feature matrix of temporal information within each sampling time point Subsequently, regarding After pooling operation, we get .

[0050] Step 2.4, , and Perform fusion and splicing to obtain the first... i Under the first working condition The moment to the The multimodal fusion vector within each time step .

[0051] Step 2.5, The input temperature / heat flux density is processed in the advanced prediction head to obtain the final result. .

[0052] Step 3, based on the first i Under the first working condition From the sampling time to the... The predicted and experimental values ​​within each sampling time point are used to construct a train electrical fire prediction network at the [number]th sampling time point. i Total loss function under each working condition Including: temperature loss function Radiative heat flux density loss function convective heat flux density loss function ; Step 3.1, construct the first equation from equation (7). i Temperature loss function under various operating conditions ; (7) In equation (7), For the first i Data loss item for temperature under each operating condition. For the first i Physical loss due to temperature under various operating conditions. These are the weighting coefficients for the temperature loss function.

[0053] Step 3.1.1, construct using equation (8) ; (8) Step 3.1.2, in this embodiment, generates in the sampling space Complementary points for a partial differential equation, Initial condition points and Considering the boundary conditions and points, neglecting the roof thickness, and taking into account two-dimensional unsteady heat conduction, for the... i Under the first working condition u Each point The temperature satisfies the two-dimensional unsteady surface energy balance equation (9): (9) In equation (9), The two-dimensional space of the roof surface. For the first i Under the first working condition u Predicted values ​​of temperature at each matching point. For the first i Under the first working condition u Predicted convective heat flux density at each alignment point For the first i Under the first working condition u Predicted radiative heat flux density at each collocation point For a two-dimensional plane Laplace operator; in this embodiment, For the equivalent surface thermal conductivity, take... , For the equivalent surface heat capacity, take , For the first u The coordinates of each collocation point in two-dimensional space. For the first u The time coordinates of each point.

[0054] Calculate the first using equations (10), (11), and (12). i Under the first working condition u PDE residuals at each collocation point temperature Boundary condition residuals and initial condition residual : (10) (11) (12) In equations (11) and (12), n For the outer normal direction of the roof, This is the Stefan-Boltzmann constant. For the first u The ambient temperature at the initial time of each coordinate point location; in this embodiment... Let be the edge equivalent convection coefficient, and take . , The edge equivalent emissivity is taken as 0.85. Take the edge equivalent ambient temperature as... .

[0055] Construct using equation (13) ; (13) In equation (13), For the first i Under the first working condition u PDE residuals at each collocation point temperature For the first i Under the first working condition u The initial condition residuals of temperature at each collocation point. For the first i Under the first working condition u Boundary condition residuals for temperature at each collocation point; , and These represent the total number of collocation points for the partial differential equation, initial conditions, and boundary conditions, respectively.

[0056] Step 3.2, construct the first equation from equation (14). i Loss function of radiative heat flux density under various operating conditions ; (14) In equation (14), For the first i Data loss term for radiative heat flux density under each operating condition. For the first i The physical loss term of radiative heat flux density under each operating condition. The weighting coefficients are for the radiative heat flux density loss function.

[0057] Step 3.2.1, construct using equation (15) ; (15) Step 3.2.2, in this embodiment, the first i Under the first working condition u Each point The radiative heat flux density satisfies the Stefan-Boltzmann law of equation (16): (16) In equation (16), For the first u The local emissivity at each point location varies little with location in actual fire scenarios. Therefore, in this embodiment, the change in local emissivity with the location of the point is not considered, and a uniform value of 0.85 is used. For the first u Equivalent radiation temperature of the local environment at each collocation point.

[0058] Calculate the first using equations (17), (18), and (19). i Under the first working condition u PDE residuals of radiative heat flux density at each collocation point Boundary condition residuals and initial condition residual : (17) (18) (19) In equations (18) and (19), express The boundary, For the first u The convective heat transfer coefficient at the initial moment at each of the coordinate points.

[0059] Construct using equation (20) ; (20) In equation (20), For the first i Under the first working condition u PDE residuals of radiative heat flux density at each collocation point For the first i Under the first working condition uInitial condition residuals of radiative heat flux density at each collocation point For the first i Under the first working condition u Boundary condition residuals of radiative heat flux density at each collocation point.

[0060] Step 3.3, construct the first equation from equation (21). i Loss function of convective heat flux density under various operating conditions ; (twenty one) In equation (21), For the first i Data loss term for convective heat flux density under each operating condition. For the first i Physical loss term of convective heat flux density under each operating condition. The weighting coefficients are for the convective heat flux density loss function.

[0061] Step 3.3.1, construct using equation (22) ; (twenty two) Step 3.3.2, in this embodiment, the first i Under the first working condition u Each point The radiative heat flux density at that point satisfies Newton's law of cooling, equation (23): (twenty three) In equation (23), For the first u Convection heat transfer coefficient at each distribution point The average flame temperature is taken as the mean temperature. In this embodiment, it is taken as... .

[0062] Calculate the first using equations (24), (25), and (26). i Under the first working condition u PDE residuals of convective heat flux density at each colligative point Boundary condition residuals and initial condition residual : (twenty four) (25) (26) Construct using equation (27) ; (27) In equation (27), For the firsti Under the first working condition u PDE residuals of convective heat flux density at each collocation point For the first i Under the first working condition u Initial condition residuals of convective heat flux density at each collocation point For the first i Under the first working condition u Boundary condition residuals of convective heat flux density at each collocation point.

[0063] Step 3.4, construct the first using equation (28). i Total loss function under each working condition : (28) In this embodiment, the weight coefficients of both the data loss term and the physical loss term are set to 1; when the physical residual dominates in the later stage of training, the weight coefficient of the physical loss term is adjusted to 0.1 to maintain the balance between the fitting accuracy of experimental data and the constraints of physical laws.

[0064] Step 4: Iteratively train the train electrical fire prediction network using the gradient descent algorithm and calculate the total loss function under all operating conditions. Adjust the network parameters until the total loss function is achieved. The process continues until convergence, resulting in a trained neural network model for predicting electrical fire situations. This model is then used to detect electrical fires from the input dataset and outputs the optimal prediction sequence.

[0065] In this embodiment, all experimental conditions are divided into training, validation, and test sets in a 7:1.5:1.5 ratio, ensuring that different combinations of train speed, heat release rate, and ventilation speed appear simultaneously in all three data subsets. The network training uses the Adam optimizer, with the initial learning rate set as follows: Momentum parameters and The parameters were set to 0.9 and 0.999 respectively, the batch size was set to 16, and the maximum number of training epochs was set to 300 epochs. Training was stopped when the total loss on the validation set did not decrease further within 20 consecutive epochs, and the network parameters with the lowest validation set loss were saved. Gradient norm pruning was used during training, with the gradient pruning threshold set to 1.0 to improve the stability of the network training process. Compared with a single-modal data-driven prediction network, this prediction network reduced the root mean square error (RMSE) of the roof temperature prediction set within the next 5-second window by 37-42%, and the mean absolute percentage error (MAPE) of the radiative heat flux density and convective heat flux density prediction sets by 32%-36% and 30%-35% respectively. The temperature prediction residual value was reduced by 58-70% compared with the pure data-driven model, and the computation inference time was approximately 140 ms. In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0066] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A method for predicting electrical fire situations of moving trains in long tunnels based on PINN, characterized in that, Includes the following steps; Step 1: Collect and construct an electrical fire dataset for long-distance tunnel moving trains using an experimental platform, including: a physical boundary dataset for long-distance tunnel moving trains. Transient flame image set of a train moving through a long tunnel. Thermal disaster field dataset of moving trains in long tunnel sections ,in, , , Indicates the first i Physical boundary datasets, transient flame image datasets, and thermal disaster field datasets under various operating conditions. N This indicates the total number of operating conditions; and , Indicates the first i Under the first working condition m Physical boundary parameters at each sampling time, Indicates the first i The total number of data collection times under each working condition; and ,in, Indicates the first i Under the first working condition m Data set of thermal disaster fields at each sampling time, , , They represent the first i Under the first working condition m The vector of experimental values ​​for roof temperature, radiative heat flux density, and convective heat flux density at each sampling time, and , , ,in, , , The first i Under the first working condition The sampling time of the first sampling moment j The temperature, the first j The radiative heat flux density and the first j Experimental values ​​of convective heat flux density at each measuring point. For the first i The total number of measuring points deployed under each working condition; right Perform fragmentation and preprocessing, the first i Atlas of Probability Cloud Map of Flame Output under Various Operating Conditions and its flame size dataset ,in, Indicates the first i Under the first working condition m Flame probability cloud map corresponding to each sampling time. Indicates the first i Under the first working condition m Flame size parameters corresponding to each sampling time; H and W for Height and width, for The number of channels; Step 2: Construct a train electrical fire prediction network, including: a physical information extraction module, a Swing Transformer flame feature extraction module, an FNO temperature / heat flux field extraction module, a multimodal information fusion module, and a temperature / heat flux density prediction head; for , , and the i Under the first working condition The moment to the Time series sequence of thermal disaster fields within a given moment Process and output the first... i Under the first working condition The moment to the The predicted value sequence of thermal disaster fields within a certain time period ,in ,in, For the time span of the predicted value sequence, , , They represent the first i Under the first working condition The predicted vector of temperature, radiative heat flux density, and convective heat flux density at each sampling time, and , , ,in, , , The first i Under the first working condition The sampling time of the first sampling moment j The temperature, the first j The radiative heat flux density and the first j Predicted values ​​for each convective heat flux density measurement point; Step 2.1, the physical information extraction module extracts information through the first fully connected layer. After performing the dimensionality upgrade operation, we obtain the first... i Under the first working condition m physical information parameter feature vector at each sampling time Then, the ReLU activation function is applied to... Perform a nonlinear transformation to obtain the first... i Under the first working condition m The eigenvectors of nonlinear parameters at each sampling time point , The input is then fed into the second fully connected layer for processing, resulting in the... i Under the first working condition m Physical feature vector at each sampling time ; Step 2.2, the Swin Transformer flame feature extraction module, includes: an image segmentation and linear embedding layer, a Swin Transformer encoder, a global feature aggregation head, and a flame geometry mapping layer, and performs... and Processing is performed to obtain the first... i Under the first working condition m Flame feature vector at each sampling time ; Step 2.3, the FNO temperature / thermal flow field extraction module includes: a linear projection layer, a temporal context stacking module, Layer FNO operator block, feature output module, and for Perform time series feature extraction, and the output includes the first... i Under the first working condition The moment to the The fusion feature vector of thermal disaster field information within a given time period ; Step 2.4, , and Perform fusion and splicing to obtain the first... i Under the first working condition The moment to the The multimodal fusion vector within each time step ; Step 2.5, The input temperature / heat flux density is processed in the advanced prediction head to obtain the final result. ; Step 3, based on the first i Under the first working condition From the sampling time to the... The predicted and experimental values ​​within each sampling time point are used to construct a train electrical fire prediction network at the [number]th sampling time point. i Total loss function under each working condition Including: temperature loss function Radiative heat flux density loss function convective heat flux density loss function ; Step 4: Iteratively train the train electrical fire prediction network using the gradient descent algorithm and calculate the total loss function under all operating conditions. Adjust the network parameters until the total loss function is reached. The process continues until convergence, resulting in a trained neural network model for predicting electrical fire situations. This model is then used to detect electrical fires from the input dataset and outputs the optimal prediction sequence.

2. The method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN according to claim 1, characterized in that, Step 2.2 includes: Step 2.2.1, the image segmentation and linear embedding layer includes: an image segmentation layer and a linear embedding layer, and is applied sequentially to... Processing is performed to obtain the first... i Under the first working condition m The initial input feature tensor at each sampling time ,in, C for The number of channels; Step 2.2.2: The Swing Transformer encoder contains four Swing Transformer base blocks and an image block merging layer between two adjacent Swing Transformer base blocks. Each Swin Transformer basic block includes, in sequence: layer normalization module, window multi-head self-attention module, feedforward neural network module, and moving window multi-head self-attention module; After four stages of feature processing, the output is... i Under the first working condition m The final output feature tensor at each sampling time point in stage 4 ; Step 2.2.3, Global Feature Aggregation Head Pair Perform global average pooling to obtain the first... i Under the first working condition m Global visual feature vector at each sampling time point ; Step 2.2.4, the flame geometry mapping layer is processed by MLP. Encode to obtain the first i Under the first working condition m Flame size feature vector at each sampling time Finally, and Vector concatenation is performed, and a non-linear mapping is applied through an MLP to finally output the first vector. i Under the first working condition m Flame feature vector at each sampling time .

3. A method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN according to claim 2, characterized in that, Step 2.2.2 includes: Step 2.2.2.1, when s When =1, let For the first i Under the first working condition m The sampling time at the ... s Input feature tensor of the stage ; Step 2.2.2.2, will Enter the first s Within the first Swing Transformer base block, and after processing by the layer normalization module, the second... s The input feature tensor after each stage of normalization ; Step 2.2.2.3, Enter the first s The window multi-head self-attention module of the basic Swing Transformer block, and the window multi-head self-attention module will... Divide into several sizes Non-overlapping local windows; Calculate the first value in any non-overlapping local window p Embedded vectors The corresponding multi-head self-attention, yielding the first p Multi-head attention fusion feature vector under each query vector ,right Perform an index rearrangement operation to obtain the first... i Under the first working condition m The sampling time at the ... s Multi-head attention fusion feature tensor of the multi-head self-attention module of the window at each stage ; then Perform a linear transformation to obtain the first... i Under the first working condition m The sampling time at the ... s Output feature tensor of the multi-head self-attention module in the phased window ; Step 2.2.2.4, the first s The feedforward neural network module in the basic Swing Transformer block uses equation (1) to... Processing is performed to obtain the first... i Under the first working condition m The sampling time at the ... s intermediate feature tensor of the stage ; (1) In equation (1), + indicates residual connection; Presentation layer normalization operation; Step 2.2.2.5, the first s The multi-head self-attention module of the moving window in the basic Swing Transformer block utilizes equation (2) to... Perform window offset and output the first... i Under the first working condition m The sampling time at the ... s Feature tensor after phase shift ; (2) In equation (2), Roll represents a cyclic shift operation. This indicates the floor function; the negative sign indicates a shift along the upper left direction. Indicates the size of the non-overlapping local window; Step 2.2.2.6, will Divide into several sizes The offset window, and calculate the first offset window within any offset window. p Embedded vectors The corresponding multi-head self-attention features are then processed in the same manner as in step 2.2.2.

3. , obtained the i Under the first working condition m The sampling time at the ... s Multi-head self-attention fusion feature tensor of the multi-head self-attention module of the moving window at each stage ; Step 2.2.2.7, using equation (3) to... Perform a reverse circular shift to obtain the first... i Under the first working condition m The sampling time at the ... s Output feature tensor of the multi-head self-attention module in the phased moving window ; (3) In equation (3), This indicates a reverse cyclic shift operation; Step 2.2.2.8, using equation (1) to... Processing is performed to obtain the first... i Under the first working condition m The sampling time at the ... s Stage output feature tensor ; Step 2.2.2.9, Enter the first s In the image patch merging layer, and Merging adjacent 2×2 image blocks in the middle yields the first... i Under the first working condition m The sampling time at the ... s Feature tensor after stage merging And perform a linear transformation to obtain the first... i Under the first working condition m The sampling time at the ... s +1 stage input feature tensor ; Step 2.2.2.10, will Assign to s Then, return to step 2.2.2.1 and execute sequentially until the output of the 4th SwingTransformer base block is obtained. .

4. A method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN according to claim 3, characterized in that, Step 2.3 includes: Step 2.3.1, the linear projection layer passes through an MLP head pair Perform a dimensionality upgrade operation to obtain the first... i Under the first working condition m Embedded feature matrix at each sampling time point Thus, the first i Under the first working condition From the sampling time to the... Embedded feature sequence within each sampling time point ; Step 2.3.2, the timing context stacking module will... Stack each element in the array and output the first element. i Under the first working condition From the sampling time to the... The fused tensor within each sampling time point ; Will After reorganizing along the time dimension, we obtain the first... i Under the first working condition From the sampling time to the... The initial input feature matrix of time series information within each sampling time point ; Step 2.3.3 Layer FNO operator block uses equation (4) for After processing, the final result is obtained. i Under the first working condition From the sampling time to the... The time sequence information within the sampling time point Layer FNO operator block output feature matrix ; (4) In equation (4), Indicates the first The output feature matrix of the -1 layer FNO operator block Indicates the first The complex weights to be learned in the layer FNO operator block. Indicates the inverse Fourier transform. Represents the Discrete Fourier Transform. Indicates a skip connection; Indicates the first The output feature matrix of the layer FNO operator block; when season ; Step 2.3.4: The feature output module first passes the feature through an MLP layer. Perform feature mapping to obtain the first... i The first working condition includes the From the sampling time to the... High-dimensional hidden layer feature matrix of temporal information within each sampling time point Subsequently, regarding Perform pooling operation to obtain .

5. A method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN according to claim 1, characterized in that, Step 3 includes: Step 3.1, construct the first equation from equation (5). i Temperature loss function under various operating conditions ; (5) In equation (5), For the first i Data loss item for temperature under each operating condition. For the first i Physical loss due to temperature under various operating conditions. These are the weighting coefficients for the temperature loss function; Step 3.2, construct the first equation from equation (6). i Loss function of radiative heat flux density under various operating conditions ; (6) In equation (6), For the first i Data loss term for radiative heat flux density under each operating condition. For the first i The physical loss term of radiative heat flux density under each operating condition. The weighting coefficients for the radiative heat flux density loss function; Step 3.3, construct the first equation from equation (7). i Loss function of convective heat flux density under various operating conditions ; (7) In equation (7), For the first i Data loss term for convective heat flux density under each operating condition. For the first i Physical loss term of convective heat flux density under each operating condition. The weighting coefficients for the convective heat flux density loss function; Step 3.4, construct the first using equation (8). i Total loss function under each working condition : (8)。 6. A method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN according to claim 5, characterized in that, Step 3.1 includes: Step 3.1.1, construct using equation (9) ; (9) Step 3.1.2: Based on the PDE residuals, initial condition residuals, and boundary condition residuals, construct using equation (10). ; (10) In equation (10), For the first i Under the first working condition u PDE residuals at each collocation point temperature For the first i Under the first working condition u The initial condition residuals of temperature at each collocation point. For the first i Under the first working condition u Boundary condition residuals for temperature at each collocation point; , and These represent the total number of collocation points for the partial differential equation, initial conditions, and boundary conditions, respectively.

7. A method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN according to claim 5, characterized in that, Step 3.2 includes: Step 3.2.1, construct using equation (11) ; (11) Step 3.2.2: Based on the PDE residuals, initial condition residuals, and boundary condition residuals, construct using equation (12). ; (12) In equation (12), For the first i Under the first working condition u PDE residuals of radiative heat flux density at each collocation point For the first i Under the first working condition u Initial condition residuals of radiative heat flux density at each collocation point For the first i Under the first working condition u Boundary condition residuals of radiative heat flux density at each collocation point.

8. A method for predicting electrical fire situations of moving trains in long-distance tunnels based on PINN according to claim 5, characterized in that, Step 3.3 includes: Step 3.3.1, construct using equation (13) ; (13) Step 3.3.2: Based on the PDE residuals, initial condition residuals, and boundary condition residuals, construct using equation (14). ; (14) In equation (14), For the first i Under the first working condition u PDE residuals of convective heat flux density at each collocation point For the first i Under the first working condition u Initial condition residuals of convective heat flux density at each collocation point For the first i Under the first working condition u Boundary condition residuals of convective heat flux density at each collocation point.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the method for predicting electrical fire situations of moving trains in long-distance tunnels according to any one of claims 1-8, and the processor is configured to execute the program stored in the memory.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When a computer program is run by a processor, it executes the steps of the method for predicting electrical fire situations of moving trains in long-distance tunnels according to any one of claims 1-8.

Images